SPL_Net

This is the code repository for our paper "SPL-Net: Spatial-Semantic Patch Learning Network for Facial Attribute Recognition with Limited Labeled Data".

Our SPL-Net is to perform FAR with limited labeled data effectively. The SPL-Net method involves a two-stage learning procedure. For the first stage, three auxiliary tasks (PRT, PST, and PCT) are jointly developed to exploit the spatial-semantic information on large-scale unlabeled facial data, and thus a powerful pretrained MSS is obtained. For the second stage, only a few number of labeled facial data are leveraged to fine-tune the pretrained MSS and an FAR model is finally learned.

Pytorch

• Python 3.7.11
• torch 1.10.1
• torchvision 0.11.2

Prepare

• The CelebA-HQ dataset is required. Random select 30 (or 300) images from CelebA-HQ, and train BiSeNet-v2 on these images to get a model.
• The CelebA, LFWA and MAAD datasets are required. Use the trained BiSeNet-v2 model to generate semantic masks of these three datasets for training PST in SPL-Net.

STAGE 1: train and test pretext task

• modify config_files/test.yml
• run python train_pretext_5b_adv.py

STAGE 2: train and test downstream task

• modify config_files/train_downstream.yml
• set PRETRAIN_EPOCH and PRETRAIN_PTH based on the saved model in stage 1
• run python train_downstream.py